Best Open Source AI

The EU AI Act and open-source models: what a company that fine-tunes one must do

The AI Act gives open-source models a partial exemption, and most companies that fine-tune one never become the provider of a new model. A reading of the articles, the Commission’s guidance and the 2026 Omnibus, ending in a checklist.

By The editors · Updated · 7 min read

A compliance officer and an engineer review a printed regulation with tabbed pages at a desk by a tall window in a European office.
For most companies, the AI Act’s duties around open models are documentation, not prohibition.
Contents
  1. The question behind the search
  2. How the Act treats open-source AI
  3. What changes when a company fine-tunes a model
  4. The dates
  5. The data protection side
  6. A checklist for a company that fine-tunes an open model
  7. Questions buyers ask
  8. Sources

A European company that moves its AI work onto an open-weight model, and then trains that model on its own data, faces a reasonable worry: whether doing so turns it into an AI model provider with the obligations that come with it. The AI Act is long, it has been amended once already, and much of the commentary on it concerns the frontier labs rather than the companies adapting their models. This guide reads the relevant articles and the Commission's guidance for that narrower case. It is a summary for orientation, not legal advice.

How the Act treats open-source AI

The Act distinguishes AI systems, the applications people use, from general-purpose AI models, the models those systems are built on⁠1.

Open-source systems. Article 2(12) says the regulation does not apply to AI systems released under free and open-source licences, unless they are placed on the market as high-risk systems, or fall under the prohibited practices of Article 5 or the transparency duties of Article 50⁠1.

Open-source models. For general-purpose models, Article 53(2) exempts providers of models released under a free and open-source licence from the technical documentation duties in Article 53(1)(a) and (b). Article 54(6) waives the requirement for non-EU providers to appoint an authorised representative. Neither exemption applies to models with systemic risk⁠1.

What still applies. Every provider of a general-purpose model, open or not, must put in place a policy to comply with EU copyright law and publish a sufficiently detailed summary of the content used for training, under Article 53(1)(c) and (d). Recital 104 confirms the open-source exemption does not cover these two duties⁠1.

What counts as open. Recital 102 ties openness to public parameters, weights, architecture and usage information. Recital 103 excludes components provided against a price or otherwise monetised, while noting that publishing through open repositories is not in itself monetisation⁠1. The Commission's guidance restates the conditions: a licence that allows access, use, modification and distribution without monetisation; public weights, architecture and usage information; and no systemic risk⁠2.

Systemic risk. A model is presumed to have systemic risk when its training used more than 10^25 floating-point operations, under Article 51(2)⁠1. The Commission's indicative test for whether a model is general-purpose at all is training compute above 10^23 operations combined with the ability to generate language, images or video⁠2.

What changes when a company fine-tunes a model

This is the point that most often worries companies, and the Commission's guidance is reassuring. A company that modifies a general-purpose model becomes the provider of the modified model only if the modification uses more than one third of the compute used to train the original model. Below that threshold, minor modifications do not trigger provider obligations⁠23.

A company above the threshold is treated as the provider of the modified model, but its documentation duties cover the modification only: what was changed and the new data used, not the whole model⁠2. Parameter-efficient fine-tuning, such as LoRA, trains a small number of added parameters⁠4, which is likely to keep a typical company fine-tune well below a third of the original model’s training compute; the figure should still be estimated and recorded.

High-risk systems follow a separate rule. Under Article 25(1), a deployer is considered the provider of a high-risk AI system if, among other cases, it puts its name or trademark on it or substantially modifies it⁠1. A company building a model into a recruitment, credit or other high-risk use case should read that article first.

The dates

DateWhat applies
2 February 2025Article 4 AI literacy duty for providers and deployers⁠5
2 August 2025Obligations for providers of general-purpose AI models⁠3
2 August 2026Commission enforcement powers over general-purpose models, including fines; national enforcement of AI literacy⁠35
27 July 2026Regulation (EU) 2026/1744, the Digital Omnibus on AI, in force⁠6
2 August 2027Deadline for general-purpose models placed on the market before 2 August 2025⁠3
2 December 2027High-risk obligations for Annex III systems, as delayed by the Omnibus⁠6
2 August 2028High-risk obligations for Annex I systems, as delayed by the Omnibus⁠6

The Omnibus, proposed by the Commission in November 2025 and adopted in July 2026, also softened Article 4, simplified documentation for small and mid-sized companies and added new prohibitions on AI that generates intimate deepfakes or child sexual abuse material⁠76. After the change, Article 4 no longer prescribes a sufficient level of AI literacy, but the Commission says that relying only on instructions for use is generally insufficient⁠5.

The voluntary General-Purpose AI Code of Practice, published in July 2025, sets out how providers can meet the transparency and copyright duties, with a separate safety and security chapter for systemic-risk models. Mistral AI and Aleph Alpha are among its signatories⁠8.

The data protection side

Training a model on company data that includes personal data is processing under the GDPR, whatever the AI Act says. The European Data Protection Board's Opinion 28/2024 is the reference: a model trained on personal data cannot in all cases be considered anonymous, and is anonymous only if both the extraction of personal data and its disclosure through queries are insignificantly likely. Legitimate interest can be a lawful basis for developing and deploying a model, subject to a three-step test, and a company deploying a third-party model should check that it was lawfully developed⁠9. France's CNIL has published practical sheets on the legal basis and on a model's status under the GDPR⁠10.

A checklist for a company that fine-tunes an open model

  1. Record each model in use, its version and its licence, and confirm whether the licence meets the Commission's open-source conditions.
  2. Check the base model's systemic-risk status. If it is above 10^25 operations, its provider carries the extra duties; a fine-tuner below the one-third threshold does not inherit them.
  3. Estimate the compute of each fine-tune and keep the figure. Below one third of the original training compute, the company is not the provider of a new general-purpose model.
  4. Document the modification anyway: what data was used, what changed and how it was evaluated. It is required above the threshold and good practice below it.
  5. Check the use case against the high-risk list and Article 25(1) before deployment.
  6. Run a GDPR assessment for the training data, minimise personal data and document the legitimate-interest test where it is the basis.
  7. Put an AI literacy programme in place for staff who use or oversee the system.
  8. Diarise the dates, in particular 2 December 2027 for high-risk systems.

A company that has done these eight things can answer the questions a regulator, customer or works council is likely to ask, and most will find that fine-tuning an open model leaves them as a deployer with documentation duties, not as a model provider.

Questions buyers ask

Does the EU AI Act apply to open-source AI?

Partly. AI systems released under free and open-source licences are outside the Act unless they are high-risk, prohibited or subject to the transparency rules. Open-source general-purpose models are exempt from some documentation duties but must still have a copyright policy and publish a summary of their training content, and models with systemic risk get no exemption.

If we fine-tune an open-source model, are we a provider under the AI Act?

Usually not. The Commission's guidance says a company that modifies a general-purpose model becomes its provider only if the modification uses more than one third of the compute used to train the original model. Even then, its obligations cover only the modification, such as the changes made and the new training data.

When do the AI Act's rules for general-purpose models apply?

The obligations for providers of general-purpose AI models have applied since 2 August 2025. The Commission's enforcement powers, including fines, apply from 2 August 2026. Models placed on the market before 2 August 2025 must comply by 2 August 2027.

Did the Digital Omnibus delay the AI Act?

It delayed the high-risk rules. Regulation (EU) 2026/1744, in force since 27 July 2026, moves the high-risk obligations to 2 December 2027 for systems listed in Annex III and to 2 August 2028 for those covered by Annex I. It also softens the AI literacy duty. It does not appear to change the obligations for general-purpose models.

What counts as a general-purpose AI model with systemic risk?

The Act presumes systemic risk when a model was trained with more than 10^25 floating-point operations of compute. Those models carry extra safety and security duties, and the open-source exemption does not apply to them.

Does the AI Act require AI training for employees?

Article 4 requires providers and deployers to take measures for AI literacy among their staff, and has applied since 2 February 2025. After the 2026 Omnibus it no longer prescribes a sufficient level, but the Commission says relying only on instructions for use is generally insufficient.

Sources

  1. Regulation (EU) 2024/1689 (Artificial Intelligence Act). EUR-Lex, 12 Jul 2024. Accessed 9 Nov 2026.
  2. Questions and answers on the guidelines for general-purpose AI providers. European Commission. Accessed 9 Nov 2026.
  3. Guidelines for providers of general-purpose AI models. European Commission. Accessed 9 Nov 2026.
  4. PEFT. Hugging Face. Accessed 9 Nov 2026.
  5. AI literacy: questions and answers. European Commission. Accessed 9 Nov 2026.
  6. Regulation (EU) 2026/1744 (Digital Omnibus on AI). EUR-Lex, 24 Jul 2026. Accessed 9 Nov 2026.
  7. Digital Omnibus on AI: regulation proposal. European Commission, 19 Nov 2025. Accessed 9 Nov 2026.
  8. The General-Purpose AI Code of Practice. European Commission, 10 Jul 2025. Accessed 9 Nov 2026.
  9. Opinion 28/2024 on certain data protection aspects related to the processing of personal data in the context of AI models. European Data Protection Board, 18 Dec 2024. Accessed 9 Nov 2026.
  10. Les fiches pratiques IA. CNIL. Accessed 9 Nov 2026.

Updated 9 Nov 2026: First published, reflecting Regulation (EU) 2026/1744 (the Digital Omnibus on AI).

Further reading

Also in: Deutsch · Français · Español · Italiano · Nederlands · Polski